[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7F9MGK024C4SSG14EEVQNVS":3,"skill-same:01M7F9MGK024C4SSG14EEVQNVS":66},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":27,"commitSha":28,"skillDirectory":29,"forksCount":30,"repositoryHtmlUrl":31,"repositoryHomepageUrl":6,"ownerHtmlUrl":32,"ownerBlogUrl":6,"ownerDescription":33,"updatedAt":34,"install":35,"resources":47},"01M7F9MGK024C4SSG14EEVQNVS","datasets",null,"Loads TUM, ICL-NUIM, and ScanNet RGB-D sequences into deterministic batched tensors, with preprocessing, pose association, semantic labels, and RGBDImages handoff guidance.","lza6\u002FArmor-piercing-system-prompts","lza6","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F128880206?v=4","01M7EB3K7WPX5T28EJ2NVHE6EG","Chinese-tingfeng","01M7EB3K97E1KQ16QWKJMFNGX8",11,"MIT","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F06-前端与设计\u002Fgradslam\u002Fsub-skills\u002Fdatasets\u002FSKILL.md",[],[],"2026-11-16 10:05:34",false,"---\nname: datasets\ndescription: \"Loads TUM, ICL-NUIM, and ScanNet RGB-D sequences into\n  deterministic batched tensors, with preprocessing, pose association, semantic\n  labels, and RGBDImages handoff guidance.\"\ndisable-model-invocation: true\nmetadata:\n  disco-role: operating\nlicense: MIT\n---\n\n# Dataset adapters operating guide\n\nUse this sub-skill when a task needs one of the repository's dataset adapters,\nits file\u002Fsequence conventions, RGB-D preprocessing, trajectory association,\nsemantic labels, or a `DataLoader` batch ready for `RGBDImages`. The adapters\nconsume data already present on disk. They do not download, extract, or repair\ndatasets.\n\nKeep the detailed contracts in the bundled references:\n\n- [API reference](references\u002Fapi-reference.md) — constructors, utility\n  functions, option gates, and output ordering.\n- [Workflows](references\u002Fworkflows.md) — preflight, sequence slicing, pose\n  normalization, and `DataLoader` handoff.\n- [Data formats](references\u002Fdata-formats.md) — directory trees, metadata line\n  layouts, units, shapes, and label encodings.\n- [Troubleshooting](references\u002Ftroubleshooting.md) — actionable diagnosis for\n  missing files, malformed associations, empty sequences, and external-data\n  limits.\n- [Layout checker](scripts\u002Fdataset_layout_check.py) — a deterministic,\n  read-only checker for user-supplied dataset paths.\n\n## Route the request\n\n1. Identify the adapter from the data contract, not from a directory name:\n   `TUM` expects timestamp lists and optional quaternion ground truth, `ICL`\n   expects TUM-compatible associations plus a block-matrix trajectory, and\n   `Scannet` expects sequence-association metadata that points into scene\n   extraction directories.\n2. Confirm the external-data boundary before constructing the dataset. The\n   package does not include these datasets, and the repository's TUM, ICL, and\n   ScanNet native tests are data-gated. A missing local root is a precondition\n   failure, not a reason to invent a substitute path or download data.\n3. Run the layout checker against paths supplied by the caller. It only reads\n   directory entries, metadata text, and selected path references; it never\n   downloads, writes, renames, deletes, or mutates input data. A nonzero result\n   means the adapter's expected structure is incomplete, not that the dataset\n   itself is corrupt.\n4. Choose `seqlen`, `dilation`, `stride`, `start`, `end`, output resolution,\n   image layout, color normalization, and return flags deliberately. Record\n   `B`, `L`, `H`, `W`, whether labels are requested, and whether poses are\n   available before handing the batch to a structure.\n\n## Construct an adapter\n\nImport directly from the dataset modules:\n\n```python\nfrom gradslam.datasets.tum import TUM\nfrom gradslam.datasets.icl import ICL\nfrom gradslam.datasets.scannet import Scannet\n```\n\nThe default return configuration asks for every output that the adapter can\nprovide. That means TUM returns color, depth, intrinsics, normalized poses,\nrelative transforms, names, and timestamps; ICL returns the same except for\nTUM timestamps; ScanNet returns those RGB-D\u002Fpose fields plus names and labels.\nReturn flags remove fields from the tuple rather than inserting `None` values,\nso unpack only the fields enabled by the configuration. See the API reference\nfor exact order.\n\nUse a tuple for an explicit selection of sequences, trajectories, or scenes.\nA string is interpreted as a split-file path only when it exists; it is not a\nsingle sequence name. `None` means all discoverable items for TUM and ICL; for\nScanNet it means all sequence metadata files. Lists are rejected by the\nadapter constructors.\n\nFor a first inspection, prefer a small `seqlen` and reduced `height`\u002F`width`,\nkeep `shuffle=False`, and request only fields needed by the next operation.\n`DataLoader` is ordinary PyTorch collation: with `batch_size=B`, each image\nfield gains a leading batch dimension and names\u002Ftimestamps are collated as a\nsequence of strings (commonly a list).\n\n## Preserve preprocessing contracts\n\n- Color is resized with bilinear interpolation. By default it remains in its\n  source-like `[0, 255]` scale; `normalize_color=True` converts it to\n  approximately `[0, 1]`.\n- Depth is resized with nearest-neighbor interpolation, gains a singleton\n  channel, and is divided by the dataset scale: `5000.0` for TUM and ICL,\n  `1000.0` for ScanNet. The resulting depth is in the adapter's meter-like\n  convention.\n- `channels_first=False` yields `(L,H,W,3)` color and `(L,H,W,1)` depth;\n  `channels_first=True` yields `(L,3,H,W)` and `(L,1,H,W)`. A collated batch\n  therefore has `(B,L,...)`. Intrinsics and poses remain `(B,1,4,4)` and\n  `(B,L,4,4)`.\n- Intrinsics are scaled independently by `height\u002F480` and `width\u002F640`.\n  Scale `fx` and `cx` by the width ratio, and `fy` and `cy` by the height\n  ratio. Do not rescale a matrix twice.\n- Pose and transform fields are only loaded when `return_pose` or\n  `return_transform` requires them. Each sequence is normalized to the first\n  pose; its first normalized pose and first relative transform are identity.\n\n## Handoff into `RGBDImages`\n\nFor an ordinary channels-last batch with all required fields:\n\n```python\nfrom torch.utils.data import DataLoader\nfrom gradslam.structures.rgbdimages import RGBDImages\n\nloader = DataLoader(dataset, batch_size=2, shuffle=False)\ncolors, depths, intrinsics, poses, transforms, names = next(iter(loader))[:6]\nrgbd = RGBDImages(colors, depths, intrinsics, poses, channels_first=False)\n```\n\nThe exact slice above is appropriate for ICL; TUM has a seventh timestamp\nfield and ScanNet has a seventh label field. The `transforms`, names,\ntimestamps, and labels are not constructor arguments for `RGBDImages`; retain\nthem separately. If `channels_first=True`, pass that same flag to\n`RGBDImages`. Do not silently omit intrinsics or replace absent poses with an\nidentity tensor: a caller that needs geometry must request\u002Fprovide them.\n\nFor ScanNet labels, keep the label tensor separate from RGB-D geometry. Labels\nare resized with nearest-neighbor interpolation and returned with a singleton\nlast channel by the implementation, so a collated batch is typically\n`(B,L,H,W,1)` even when `channels_first=True`; older prose may describe it as\n`(L,H,W)`. Use `seg_classes=\"nyu40\"` to retain the source indexing, or\n`seg_classes=\"scannet20\"` to remap the supported source ids to the contiguous\n20-class palette. Use `get_color_encoding` and\n`datautils.create_label_image` only when a visualization or colorized label\nimage is explicitly needed, and verify the palette representation against the\ninstalled release.\n\n## Boundaries and verification\n\nThis skill does not promise that any external dataset is present, that a split\nfile names valid items, or that all referenced image\u002Fpose files are readable.\nThe checker reports path existence only; adapter construction and one small\n`__getitem__` are the next data-dependent verification step. Do not run\nnetwork downloads, full native dataset tests, GPU flows, GUI visualization, or\nnotebook execution while drafting or when only a layout check was requested.\n\nThe deterministic CPU utility coverage is the strongest data-independent\nsignal. TUM, ICL, and ScanNet adapter tests\u002Fexamples are **not guaranteed**\nwithout their external data and prepared extraction metadata. Keep that status\nexplicit in reports and handoffs.\n","# Dataset adapters operating guide\n\nUse this sub-skill when a task needs one of the repository's dataset adapters,\nits file\u002Fsequence conventions, RGB-D preprocessing, trajectory association,\nsemantic labels, or a `DataLoader` batch ready for `RGBDImages`. The adapters\nconsume data already present on disk. They do not download, extract, or repair\ndatasets.\n\nKeep the detailed contracts in the bundled references:\n\n- [API reference](references\u002Fapi-reference.md) — constructors, utility\n  functions, option gates, and output ordering.\n- [Workflows](references\u002Fworkflows.md) — preflight, sequence slicing, pose\n  normalization, and `DataLoader` handoff.\n- [Data formats](references\u002Fdata-formats.md) — directory trees, metadata line\n  layouts, units, shapes, and label encodings.\n- [Troubleshooting](references\u002Ftroubleshooting.md) — actionable diagnosis for\n  missing files, malformed associations, empty sequences, and external-data\n  limits.\n- [Layout checker](scripts\u002Fdataset_layout_check.py) — a deterministic,\n  read-only checker for user-supplied dataset paths.\n\n## Route the request\n\n1. Identify the adapter from the data contract, not from a directory name:\n   `TUM` expects timestamp lists and optional quaternion ground truth, `ICL`\n   expects TUM-compatible associations plus a block-matrix trajectory, and\n   `Scannet` expects sequence-association metadata that points into scene\n   extraction directories.\n2. Confirm the external-data boundary before constructing the dataset. The\n   package does not include these datasets, and the repository's TUM, ICL, and\n   ScanNet native tests are data-gated. A missing local root is a precondition\n   failure, not a reason to invent a substitute path or download data.\n3. Run the layout checker against paths supplied by the caller. It only reads\n   directory entries, metadata text, and selected path references; it never\n   downloads, writes, renames, deletes, or mutates input data. A nonzero result\n   means the adapter's expected structure is incomplete, not that the dataset\n   itself is corrupt.\n4. Choose `seqlen`, `dilation`, `stride`, `start`, `end`, output resolution,\n   image layout, color normalization, and return flags deliberately. Record\n   `B`, `L`, `H`, `W`, whether labels are requested, and whether poses are\n   available before handing the batch to a structure.\n\n## Construct an adapter\n\nImport directly from the dataset modules:\n\n```python\nfrom gradslam.datasets.tum import TUM\nfrom gradslam.datasets.icl import ICL\nfrom gradslam.datasets.scannet import Scannet\n```\n\nThe default return configuration asks for every output that the adapter can\nprovide. That means TUM returns color, depth, intrinsics, normalized poses,\nrelative transforms, names, and timestamps; ICL returns the same except for\nTUM timestamps; ScanNet returns those RGB-D\u002Fpose fields plus names and labels.\nReturn flags remove fields from the tuple rather than inserting `None` values,\nso unpack only the fields enabled by the configuration. See the API reference\nfor exact order.\n\nUse a tuple for an explicit selection of sequences, trajectories, or scenes.\nA string is interpreted as a split-file path only when it exists; it is not a\nsingle sequence name. `None` means all discoverable items for TUM and ICL; for\nScanNet it means all sequence metadata files. Lists are rejected by the\nadapter constructors.\n\nFor a first inspection, prefer a small `seqlen` and reduced `height`\u002F`width`,\nkeep `shuffle=False`, and request only fields needed by the next operation.\n`DataLoader` is ordinary PyTorch collation: with `batch_size=B`, each image\nfield gains a leading batch dimension and names\u002Ftimestamps are collated as a\nsequence of strings (commonly a list).\n\n## Preserve preprocessing contracts\n\n- Color is resized with bilinear interpolation. By default it remains in its\n  source-like `[0, 255]` scale; `normalize_color=True` converts it to\n  approximately `[0, 1]`.\n- Depth is resized with nearest-neighbor interpolation, gains a singleton\n  channel, and is divided by the dataset scale: `5000.0` for TUM and ICL,\n  `1000.0` for ScanNet. The resulting depth is in the adapter's meter-like\n  convention.\n- `channels_first=False` yields `(L,H,W,3)` color and `(L,H,W,1)` depth;\n  `channels_first=True` yields `(L,3,H,W)` and `(L,1,H,W)`. A collated batch\n  therefore has `(B,L,...)`. Intrinsics and poses remain `(B,1,4,4)` and\n  `(B,L,4,4)`.\n- Intrinsics are scaled independently by `height\u002F480` and `width\u002F640`.\n  Scale `fx` and `cx` by the width ratio, and `fy` and `cy` by the height\n  ratio. Do not rescale a matrix twice.\n- Pose and transform fields are only loaded when `return_pose` or\n  `return_transform` requires them. Each sequence is normalized to the first\n  pose; its first normalized pose and first relative transform are identity.\n\n## Handoff into `RGBDImages`\n\nFor an ordinary channels-last batch with all required fields:\n\n```python\nfrom torch.utils.data import DataLoader\nfrom gradslam.structures.rgbdimages import RGBDImages\n\nloader = DataLoader(dataset, batch_size=2, shuffle=False)\ncolors, depths, intrinsics, poses, transforms, names = next(iter(loader))[:6]\nrgbd = RGBDImages(colors, depths, intrinsics, poses, channels_first=False)\n```\n\nThe exact slice above is appropriate for ICL; TUM has a seventh timestamp\nfield and ScanNet has a seventh label field. The `transforms`, names,\ntimestamps, and labels are not constructor arguments for `RGBDImages`; retain\nthem separately. If `channels_first=True`, pass that same flag to\n`RGBDImages`. Do not silently omit intrinsics or replace absent poses with an\nidentity tensor: a caller that needs geometry must request\u002Fprovide them.\n\nFor ScanNet labels, keep the label tensor separate from RGB-D geometry. Labels\nare resized with nearest-neighbor interpolation and returned with a singleton\nlast channel by the implementation, so a collated batch is typically\n`(B,L,H,W,1)` even when `channels_first=True`; older prose may describe it as\n`(L,H,W)`. Use `seg_classes=\"nyu40\"` to retain the source indexing, or\n`seg_classes=\"scannet20\"` to remap the supported source ids to the contiguous\n20-class palette. Use `get_color_encoding` and\n`datautils.create_label_image` only when a visualization or colorized label\nimage is explicitly needed, and verify the palette representation against the\ninstalled release.\n\n## Boundaries and verification\n\nThis skill does not promise that any external dataset is present, that a split\nfile names valid items, or that all referenced image\u002Fpose files are readable.\nThe checker reports path existence only; adapter construction and one small\n`__getitem__` are the next data-dependent verification step. Do not run\nnetwork downloads, full native dataset tests, GPU flows, GUI visualization, or\nnotebook execution while drafting or when only a layout check was requested.\n\nThe deterministic CPU utility coverage is the strongest data-independent\nsignal. TUM, ICL, and ScanNet adapter tests\u002Fexamples are **not guaranteed**\nwithout their external data and prepared extraction metadata. Keep that status\nexplicit in reports and handoffs.\n",{"name":5,"license":15,"metadata":24,"description":7,"disable-model-invocation":26},{"disco-role":25},"operating",true,"main","d69096e239f2a89aabc2cd078c47664a9d3b1e11","06-前端与设计\u002Fgradslam\u002Fsub-skills\u002Fdatasets",3,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts","https:\u002F\u002Fgithub.com\u002Flza6","个人vx：Tf00798\r\n个人✈：@TFGY999","2026-10-08 14:29:22",{"source":8,"skillName":5,"commands":36,"prompt":46,"zipAvailable":26},[37,40,43],{"tool":38,"command":39},"npx","npx skills add lza6\u002FArmor-piercing-system-prompts --skill datasets",{"tool":41,"command":42},"bunx","bunx skills add lza6\u002FArmor-piercing-system-prompts --skill datasets",{"tool":44,"command":45},"pnpm","pnpm dlx skills add lza6\u002FArmor-piercing-system-prompts --skill datasets","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F06-前端与设计\u002Fgradslam\u002Fsub-skills\u002Fdatasets\u002FSKILL.md",[48,54,58,62],{"path":49,"resourceType":50,"resourceTypeDesc":51,"sizeBytes":52,"sourceHtmlUrl":53},"references\u002Fapi-reference.md","REFERENCE","参考文档",7769,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F06-前端与设计\u002Fgradslam\u002Fsub-skills\u002Fdatasets\u002Freferences\u002Fapi-reference.md",{"path":55,"resourceType":50,"resourceTypeDesc":51,"sizeBytes":56,"sourceHtmlUrl":57},"references\u002Fdata-formats.md",4614,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F06-前端与设计\u002Fgradslam\u002Fsub-skills\u002Fdatasets\u002Freferences\u002Fdata-formats.md",{"path":59,"resourceType":50,"resourceTypeDesc":51,"sizeBytes":60,"sourceHtmlUrl":61},"references\u002Ftroubleshooting.md",7338,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F06-前端与设计\u002Fgradslam\u002Fsub-skills\u002Fdatasets\u002Freferences\u002Ftroubleshooting.md",{"path":63,"resourceType":50,"resourceTypeDesc":51,"sizeBytes":64,"sourceHtmlUrl":65},"references\u002Fworkflows.md",7140,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F06-前端与设计\u002Fgradslam\u002Fsub-skills\u002Fdatasets\u002Freferences\u002Fworkflows.md",[67,74,79,86,93,100],{"skillNo":68,"name":69,"nameZh":6,"description":70,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":71,"categories":72,"keywords":73,"publishedAt":19},"01M7FFG485KY83TD6HSC4YPCMC","obsidian-clipper-template-creator","Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobsidian-clipper-template-creator\u002FSKILL.md",[],[],{"skillNo":75,"name":69,"nameZh":6,"description":70,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":76,"categories":77,"keywords":78,"publishedAt":19},"01M7FFG0T6AZ0STEJPSB3CX84K","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobsidian-clipper-template-creator-antigravity-awesome-skills-main\u002FSKILL.md",[],[],{"skillNo":80,"name":81,"nameZh":6,"description":82,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":83,"categories":84,"keywords":85,"publishedAt":19},"01M7FFFY0QEM14SKWPDYGMXM2P","python-api","Programmatic Object-Detection-Metrics API guidance for BoundingBox, BoundingBoxes, Evaluator, enums, coordinate conversion, AP\u002FmAP semantics, and source-style import diagnostics.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobject-detection-metrics\u002Fsub-skills\u002Fpython-api\u002FSKILL.md",[],[],{"skillNo":87,"name":88,"nameZh":6,"description":89,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":90,"categories":91,"keywords":92,"publishedAt":19},"01M7FFFVZQZ2GJ0C0TGP7JHD20","file-evaluation","Folder-based PASCAL VOC AP\u002FmAP evaluation from ground-truth and detection text files.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobject-detection-metrics\u002Fsub-skills\u002Ffile-evaluation\u002FSKILL.md",[],[],{"skillNo":94,"name":95,"nameZh":6,"description":96,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":97,"categories":98,"keywords":99,"publishedAt":19},"01M7FFFSFNYSE77SJ6PE1CK2DA","object-detection-metrics","Guides legacy Object-Detection-Metrics PASCAL VOC AP\u002FmAP evaluation from detection text folders and direct Python API classes.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobject-detection-metrics\u002FSKILL.md",[],[],{"skillNo":101,"name":102,"nameZh":6,"description":103,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":104,"categories":105,"keywords":106,"publishedAt":19},"01M7FFFP0GCK1RVXBCPQ8APBMT","nx-workspace-patterns","Configure and optimize Nx monorepo workspaces. Use when setting up Nx, configuring project boundaries, optimizing build caching, or implementing affected commands.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fnx-workspace-patterns\u002FSKILL.md",[],[]]